sys/kern/sys_pipe.c

   1 /*-
   2  * Copyright (c) 1996 John S. Dyson
   3  * Copyright (c) 2012 Giovanni Trematerra
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice immediately at the beginning of the file, without modification,
  11  *    this list of conditions, and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Absolutely no warranty of function or purpose is made by the author
  16  *    John S. Dyson.
  17  * 4. Modifications may be freely made to this file if the above conditions
  18  *    are met.
  19  */
  20
  21 /*
  22  * This file contains a high-performance replacement for the socket-based
  23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  24  * all features of sockets, but does do everything that pipes normally
  25  * do.
  26  */
  27
  28 /*
  29  * This code has two modes of operation, a small write mode and a large
  30  * write mode.  The small write mode acts like conventional pipes with
  31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  34  * memory, and the receiving process copies directly from these pinned pages
  35  * in the sending process.
  36  *
  37  * If the sending process receives a signal, it is possible that it will
  38  * go away, and certainly its address space can change, because control
  39  * is returned back to the user-mode side.  In that case, the pipe code
  40  * arranges to copy the buffer supplied by the user process, to a pageable
  41  * kernel buffer, and the receiving process will grab the data from the
  42  * pageable kernel buffer.  Since signals don't happen all that often,
  43  * the copy operation is normally eliminated.
  44  *
  45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  46  * happen for small transfers so that the system will not spend all of
  47  * its time context switching.
  48  *
  49  * In order to limit the resource use of pipes, two sysctls exist:
  50  *
  51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  52  * address space available to us in pipe_map. This value is normally
  53  * autotuned, but may also be loader tuned.
  54  *
  55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  56  * memory in use by pipes.
  57  *
  58  * Based on how large pipekva is relative to maxpipekva, the following
  59  * will happen:
  60  *
  61  * 0% - 50%:
  62  *     New pipes are given 16K of memory backing, pipes may dynamically
  63  *     grow to as large as 64K where needed.
  64  * 50% - 75%:
  65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  66  *     existing pipes may NOT grow.
  67  * 75% - 100%:
  68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  69  *     existing pipes will be shrunk down to 4K whenever possible.
  70  *
  71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  73  * resize which MUST occur for reverse-direction pipes when they are
  74  * first used.
  75  *
  76  * Additional information about the current state of pipes may be obtained
  77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  78  * and kern.ipc.piperesizefail.
  79  *
  80  * Locking rules:  There are two locks present here:  A mutex, used via
  81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  82  * the flag, as mutexes can not persist over uiomove.  The mutex
  83  * exists only to guard access to the flag, and is not in itself a
  84  * locking mechanism.  Also note that there is only a single mutex for
  85  * both directions of a pipe.
  86  *
  87  * As pipelock() may have to sleep before it can acquire the flag, it
  88  * is important to reread all data after a call to pipelock(); everything
  89  * in the structure may have changed.
  90  */
  91
  92 #include <sys/cdefs.h>
  93 __FBSDID("$FreeBSD$");
  94
  95 #include <sys/param.h>
  96 #include <sys/systm.h>
  97 #include <sys/conf.h>
  98 #include <sys/fcntl.h>
  99 #include <sys/file.h>
 100 #include <sys/filedesc.h>
 101 #include <sys/filio.h>
 102 #include <sys/kernel.h>
 103 #include <sys/lock.h>
 104 #include <sys/mutex.h>
 105 #include <sys/ttycom.h>
 106 #include <sys/stat.h>
 107 #include <sys/malloc.h>
 108 #include <sys/poll.h>
 109 #include <sys/selinfo.h>
 110 #include <sys/signalvar.h>
 111 #include <sys/syscallsubr.h>
 112 #include <sys/sysctl.h>
 113 #include <sys/sysproto.h>
 114 #include <sys/pipe.h>
 115 #include <sys/proc.h>
 116 #include <sys/vnode.h>
 117 #include <sys/uio.h>
 118 #include <sys/event.h>
 119
 120 #include <security/mac/mac_framework.h>
 121
 122 #include <vm/vm.h>
 123 #include <vm/vm_param.h>
 124 #include <vm/vm_object.h>
 125 #include <vm/vm_kern.h>
 126 #include <vm/vm_extern.h>
 127 #include <vm/pmap.h>
 128 #include <vm/vm_map.h>
 129 #include <vm/vm_page.h>
 130 #include <vm/uma.h>
 131
 132 /*
 133  * Use this define if you want to disable *fancy* VM things.  Expect an
 134  * approx 30% decrease in transfer rate.  This could be useful for
 135  * NetBSD or OpenBSD.
 136  */
 137 /* #define PIPE_NODIRECT */
 138
 139 #define PIPE_PEER(pipe) \
 140         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 141
 142 /*
 143  * interfaces to the outside world
 144  */
 145 static fo_rdwr_t        pipe_read;
 146 static fo_rdwr_t        pipe_write;
 147 static fo_truncate_t    pipe_truncate;
 148 static fo_ioctl_t       pipe_ioctl;
 149 static fo_poll_t        pipe_poll;
 150 static fo_kqfilter_t    pipe_kqfilter;
 151 static fo_stat_t        pipe_stat;
 152 static fo_close_t       pipe_close;
 153 static fo_chmod_t       pipe_chmod;
 154 static fo_chown_t       pipe_chown;
 155
 156 struct fileops pipeops = {
 157         .fo_read = pipe_read,
 158         .fo_write = pipe_write,
 159         .fo_truncate = pipe_truncate,
 160         .fo_ioctl = pipe_ioctl,
 161         .fo_poll = pipe_poll,
 162         .fo_kqfilter = pipe_kqfilter,
 163         .fo_stat = pipe_stat,
 164         .fo_close = pipe_close,
 165         .fo_chmod = pipe_chmod,
 166         .fo_chown = pipe_chown,
 167         .fo_sendfile = invfo_sendfile,
 168         .fo_flags = DFLAG_PASSABLE
 169 };
 170
 171 static void     filt_pipedetach(struct knote *kn);
 172 static void     filt_pipedetach_notsup(struct knote *kn);
 173 static int      filt_pipenotsup(struct knote *kn, long hint);
 174 static int      filt_piperead(struct knote *kn, long hint);
 175 static int      filt_pipewrite(struct knote *kn, long hint);
 176
 177 static struct filterops pipe_nfiltops = {
 178         .f_isfd = 1,
 179         .f_detach = filt_pipedetach_notsup,
 180         .f_event = filt_pipenotsup
 181 };
 182 static struct filterops pipe_rfiltops = {
 183         .f_isfd = 1,
 184         .f_detach = filt_pipedetach,
 185         .f_event = filt_piperead
 186 };
 187 static struct filterops pipe_wfiltops = {
 188         .f_isfd = 1,
 189         .f_detach = filt_pipedetach,
 190         .f_event = filt_pipewrite
 191 };
 192
 193 /*
 194  * Default pipe buffer size(s), this can be kind-of large now because pipe
 195  * space is pageable.  The pipe code will try to maintain locality of
 196  * reference for performance reasons, so small amounts of outstanding I/O
 197  * will not wipe the cache.
 198  */
 199 #define MINPIPESIZE (PIPE_SIZE/3)
 200 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 201
 202 static long amountpipekva;
 203 static int pipefragretry;
 204 static int pipeallocfail;
 205 static int piperesizefail;
 206 static int piperesizeallowed = 1;
 207
 208 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 209            &maxpipekva, 0, "Pipe KVA limit");
 210 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 211            &amountpipekva, 0, "Pipe KVA usage");
 212 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 213           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 214 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 215           &pipeallocfail, 0, "Pipe allocation failures");
 216 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 217           &piperesizefail, 0, "Pipe resize failures");
 218 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 219           &piperesizeallowed, 0, "Pipe resizing allowed");
 220
 221 static void pipeinit(void *dummy __unused);
 222 static void pipeclose(struct pipe *cpipe);
 223 static void pipe_free_kmem(struct pipe *cpipe);
 224 static int pipe_create(struct pipe *pipe, int backing);
 225 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 226 static __inline int pipelock(struct pipe *cpipe, int catch);
 227 static __inline void pipeunlock(struct pipe *cpipe);
 228 #ifndef PIPE_NODIRECT
 229 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 230 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 231 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 232 static void pipe_clone_write_buffer(struct pipe *wpipe);
 233 #endif
 234 static int pipespace(struct pipe *cpipe, int size);
 235 static int pipespace_new(struct pipe *cpipe, int size);
 236
 237 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 238 static int      pipe_zone_init(void *mem, int size, int flags);
 239 static void     pipe_zone_fini(void *mem, int size);
 240
 241 static uma_zone_t pipe_zone;
 242 static struct unrhdr *pipeino_unr;
 243 static dev_t pipedev_ino;
 244
 245 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 246
 247 static void
 248 pipeinit(void *dummy __unused)
 249 {
 250
 251         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 252             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 253             UMA_ALIGN_PTR, 0);
 254         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 255         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 256         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 257         pipedev_ino = devfs_alloc_cdp_inode();
 258         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 259 }
 260
 261 static int
 262 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 263 {
 264         struct pipepair *pp;
 265         struct pipe *rpipe, *wpipe;
 266
 267         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 268
 269         pp = (struct pipepair *)mem;
 270
 271         /*
 272          * We zero both pipe endpoints to make sure all the kmem pointers
 273          * are NULL, flag fields are zero'd, etc.  We timestamp both
 274          * endpoints with the same time.
 275          */
 276         rpipe = &pp->pp_rpipe;
 277         bzero(rpipe, sizeof(*rpipe));
 278         vfs_timestamp(&rpipe->pipe_ctime);
 279         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 280
 281         wpipe = &pp->pp_wpipe;
 282         bzero(wpipe, sizeof(*wpipe));
 283         wpipe->pipe_ctime = rpipe->pipe_ctime;
 284         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 285
 286         rpipe->pipe_peer = wpipe;
 287         rpipe->pipe_pair = pp;
 288         wpipe->pipe_peer = rpipe;
 289         wpipe->pipe_pair = pp;
 290
 291         /*
 292          * Mark both endpoints as present; they will later get free'd
 293          * one at a time.  When both are free'd, then the whole pair
 294          * is released.
 295          */
 296         rpipe->pipe_present = PIPE_ACTIVE;
 297         wpipe->pipe_present = PIPE_ACTIVE;
 298
 299         /*
 300          * Eventually, the MAC Framework may initialize the label
 301          * in ctor or init, but for now we do it elswhere to avoid
 302          * blocking in ctor or init.
 303          */
 304         pp->pp_label = NULL;
 305
 306         return (0);
 307 }
 308
 309 static int
 310 pipe_zone_init(void *mem, int size, int flags)
 311 {
 312         struct pipepair *pp;
 313
 314         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 315
 316         pp = (struct pipepair *)mem;
 317
 318         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
 319         return (0);
 320 }
 321
 322 static void
 323 pipe_zone_fini(void *mem, int size)
 324 {
 325         struct pipepair *pp;
 326
 327         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 328
 329         pp = (struct pipepair *)mem;
 330
 331         mtx_destroy(&pp->pp_mtx);
 332 }
 333
 334 static int
 335 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 336 {
 337         struct pipepair *pp;
 338         struct pipe *rpipe, *wpipe;
 339         int error;
 340
 341         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 342 #ifdef MAC
 343         /*
 344          * The MAC label is shared between the connected endpoints.  As a
 345          * result mac_pipe_init() and mac_pipe_create() are called once
 346          * for the pair, and not on the endpoints.
 347          */
 348         mac_pipe_init(pp);
 349         mac_pipe_create(td->td_ucred, pp);
 350 #endif
 351         rpipe = &pp->pp_rpipe;
 352         wpipe = &pp->pp_wpipe;
 353
 354         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 355         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 356
 357         /* Only the forward direction pipe is backed by default */
 358         if ((error = pipe_create(rpipe, 1)) != 0 ||
 359             (error = pipe_create(wpipe, 0)) != 0) {
 360                 pipeclose(rpipe);
 361                 pipeclose(wpipe);
 362                 return (error);
 363         }
 364
 365         rpipe->pipe_state |= PIPE_DIRECTOK;
 366         wpipe->pipe_state |= PIPE_DIRECTOK;
 367         return (0);
 368 }
 369
 370 int
 371 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 372 {
 373         struct pipepair *pp;
 374         int error;
 375
 376         error = pipe_paircreate(td, &pp);
 377         if (error != 0)
 378                 return (error);
 379         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 380         *ppipe = &pp->pp_rpipe;
 381         return (0);
 382 }
 383
 384 void
 385 pipe_dtor(struct pipe *dpipe)
 386 {
 387         ino_t ino;
 388
 389         ino = dpipe->pipe_ino;
 390         funsetown(&dpipe->pipe_sigio);
 391         pipeclose(dpipe);
 392         if (dpipe->pipe_state & PIPE_NAMED) {
 393                 dpipe = dpipe->pipe_peer;
 394                 funsetown(&dpipe->pipe_sigio);
 395                 pipeclose(dpipe);
 396         }
 397         if (ino != 0 && ino != (ino_t)-1)
 398                 free_unr(pipeino_unr, ino);
 399 }
 400
 401 /*
 402  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 403  * the zone pick up the pieces via pipeclose().
 404  */
 405 int
 406 kern_pipe(struct thread *td, int fildes[2])
 407 {
 408
 409         return (kern_pipe2(td, fildes, 0));
 410 }
 411
 412 int
 413 kern_pipe2(struct thread *td, int fildes[2], int flags)
 414 {
 415         struct filedesc *fdp;
 416         struct file *rf, *wf;
 417         struct pipe *rpipe, *wpipe;
 418         struct pipepair *pp;
 419         int fd, fflags, error;
 420
 421         fdp = td->td_proc->p_fd;
 422         error = pipe_paircreate(td, &pp);
 423         if (error != 0)
 424                 return (error);
 425         rpipe = &pp->pp_rpipe;
 426         wpipe = &pp->pp_wpipe;
 427         error = falloc(td, &rf, &fd, flags);
 428         if (error) {
 429                 pipeclose(rpipe);
 430                 pipeclose(wpipe);
 431                 return (error);
 432         }
 433         /* An extra reference on `rf' has been held for us by falloc(). */
 434         fildes[0] = fd;
 435
 436         fflags = FREAD | FWRITE;
 437         if ((flags & O_NONBLOCK) != 0)
 438                 fflags |= FNONBLOCK;
 439
 440         /*
 441          * Warning: once we've gotten past allocation of the fd for the
 442          * read-side, we can only drop the read side via fdrop() in order
 443          * to avoid races against processes which manage to dup() the read
 444          * side while we are blocked trying to allocate the write side.
 445          */
 446         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 447         error = falloc(td, &wf, &fd, flags);
 448         if (error) {
 449                 fdclose(fdp, rf, fildes[0], td);
 450                 fdrop(rf, td);
 451                 /* rpipe has been closed by fdrop(). */
 452                 pipeclose(wpipe);
 453                 return (error);
 454         }
 455         /* An extra reference on `wf' has been held for us by falloc(). */
 456         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 457         fdrop(wf, td);
 458         fildes[1] = fd;
 459         fdrop(rf, td);
 460
 461         return (0);
 462 }
 463
 464 /* ARGSUSED */
 465 int
 466 sys_pipe(struct thread *td, struct pipe_args *uap)
 467 {
 468         int error;
 469         int fildes[2];
 470
 471         error = kern_pipe(td, fildes);
 472         if (error)
 473                 return (error);
 474
 475         td->td_retval[0] = fildes[0];
 476         td->td_retval[1] = fildes[1];
 477
 478         return (0);
 479 }
 480
 481 int
 482 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 483 {
 484         int error, fildes[2];
 485
 486         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 487                 return (EINVAL);
 488         error = kern_pipe2(td, fildes, uap->flags);
 489         if (error)
 490                 return (error);
 491         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 492         if (error) {
 493                 (void)kern_close(td, fildes[0]);
 494                 (void)kern_close(td, fildes[1]);
 495         }
 496         return (error);
 497 }
 498
 499 /*
 500  * Allocate kva for pipe circular buffer, the space is pageable
 501  * This routine will 'realloc' the size of a pipe safely, if it fails
 502  * it will retain the old buffer.
 503  * If it fails it will return ENOMEM.
 504  */
 505 static int
 506 pipespace_new(cpipe, size)
 507         struct pipe *cpipe;
 508         int size;
 509 {
 510         caddr_t buffer;
 511         int error, cnt, firstseg;
 512         static int curfail = 0;
 513         static struct timeval lastfail;
 514
 515         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 516         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 517                 ("pipespace: resize of direct writes not allowed"));
 518 retry:
 519         cnt = cpipe->pipe_buffer.cnt;
 520         if (cnt > size)
 521                 size = cnt;
 522
 523         size = round_page(size);
 524         buffer = (caddr_t) vm_map_min(pipe_map);
 525
 526         error = vm_map_find(pipe_map, NULL, 0,
 527                 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
 528                 VM_PROT_ALL, VM_PROT_ALL, 0);
 529         if (error != KERN_SUCCESS) {
 530                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 531                         (size > SMALL_PIPE_SIZE)) {
 532                         size = SMALL_PIPE_SIZE;
 533                         pipefragretry++;
 534                         goto retry;
 535                 }
 536                 if (cpipe->pipe_buffer.buffer == NULL) {
 537                         pipeallocfail++;
 538                         if (ppsratecheck(&lastfail, &curfail, 1))
 539                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 540                 } else {
 541                         piperesizefail++;
 542                 }
 543                 return (ENOMEM);
 544         }
 545
 546         /* copy data, then free old resources if we're resizing */
 547         if (cnt > 0) {
 548                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 549                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 550                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 551                                 buffer, firstseg);
 552                         if ((cnt - firstseg) > 0)
 553                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 554                                         cpipe->pipe_buffer.in);
 555                 } else {
 556                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 557                                 buffer, cnt);
 558                 }
 559         }
 560         pipe_free_kmem(cpipe);
 561         cpipe->pipe_buffer.buffer = buffer;
 562         cpipe->pipe_buffer.size = size;
 563         cpipe->pipe_buffer.in = cnt;
 564         cpipe->pipe_buffer.out = 0;
 565         cpipe->pipe_buffer.cnt = cnt;
 566         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 567         return (0);
 568 }
 569
 570 /*
 571  * Wrapper for pipespace_new() that performs locking assertions.
 572  */
 573 static int
 574 pipespace(cpipe, size)
 575         struct pipe *cpipe;
 576         int size;
 577 {
 578
 579         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 580                 ("Unlocked pipe passed to pipespace"));
 581         return (pipespace_new(cpipe, size));
 582 }
 583
 584 /*
 585  * lock a pipe for I/O, blocking other access
 586  */
 587 static __inline int
 588 pipelock(cpipe, catch)
 589         struct pipe *cpipe;
 590         int catch;
 591 {
 592         int error;
 593
 594         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 595         while (cpipe->pipe_state & PIPE_LOCKFL) {
 596                 cpipe->pipe_state |= PIPE_LWANT;
 597                 error = msleep(cpipe, PIPE_MTX(cpipe),
 598                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 599                     "pipelk", 0);
 600                 if (error != 0)
 601                         return (error);
 602         }
 603         cpipe->pipe_state |= PIPE_LOCKFL;
 604         return (0);
 605 }
 606
 607 /*
 608  * unlock a pipe I/O lock
 609  */
 610 static __inline void
 611 pipeunlock(cpipe)
 612         struct pipe *cpipe;
 613 {
 614
 615         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 616         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 617                 ("Unlocked pipe passed to pipeunlock"));
 618         cpipe->pipe_state &= ~PIPE_LOCKFL;
 619         if (cpipe->pipe_state & PIPE_LWANT) {
 620                 cpipe->pipe_state &= ~PIPE_LWANT;
 621                 wakeup(cpipe);
 622         }
 623 }
 624
 625 void
 626 pipeselwakeup(cpipe)
 627         struct pipe *cpipe;
 628 {
 629
 630         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 631         if (cpipe->pipe_state & PIPE_SEL) {
 632                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 633                 if (!SEL_WAITING(&cpipe->pipe_sel))
 634                         cpipe->pipe_state &= ~PIPE_SEL;
 635         }
 636         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 637                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 638         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 639 }
 640
 641 /*
 642  * Initialize and allocate VM and memory for pipe.  The structure
 643  * will start out zero'd from the ctor, so we just manage the kmem.
 644  */
 645 static int
 646 pipe_create(pipe, backing)
 647         struct pipe *pipe;
 648         int backing;
 649 {
 650         int error;
 651
 652         if (backing) {
 653                 if (amountpipekva > maxpipekva / 2)
 654                         error = pipespace_new(pipe, SMALL_PIPE_SIZE);
 655                 else
 656                         error = pipespace_new(pipe, PIPE_SIZE);
 657         } else {
 658                 /* If we're not backing this pipe, no need to do anything. */
 659                 error = 0;
 660         }
 661         pipe->pipe_ino = -1;
 662         return (error);
 663 }
 664
 665 /* ARGSUSED */
 666 static int
 667 pipe_read(fp, uio, active_cred, flags, td)
 668         struct file *fp;
 669         struct uio *uio;
 670         struct ucred *active_cred;
 671         struct thread *td;
 672         int flags;
 673 {
 674         struct pipe *rpipe;
 675         int error;
 676         int nread = 0;
 677         int size;
 678
 679         rpipe = fp->f_data;
 680         PIPE_LOCK(rpipe);
 681         ++rpipe->pipe_busy;
 682         error = pipelock(rpipe, 1);
 683         if (error)
 684                 goto unlocked_error;
 685
 686 #ifdef MAC
 687         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 688         if (error)
 689                 goto locked_error;
 690 #endif
 691         if (amountpipekva > (3 * maxpipekva) / 4) {
 692                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 693                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 694                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 695                         (piperesizeallowed == 1)) {
 696                         PIPE_UNLOCK(rpipe);
 697                         pipespace(rpipe, SMALL_PIPE_SIZE);
 698                         PIPE_LOCK(rpipe);
 699                 }
 700         }
 701
 702         while (uio->uio_resid) {
 703                 /*
 704                  * normal pipe buffer receive
 705                  */
 706                 if (rpipe->pipe_buffer.cnt > 0) {
 707                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 708                         if (size > rpipe->pipe_buffer.cnt)
 709                                 size = rpipe->pipe_buffer.cnt;
 710                         if (size > uio->uio_resid)
 711                                 size = uio->uio_resid;
 712
 713                         PIPE_UNLOCK(rpipe);
 714                         error = uiomove(
 715                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 716                             size, uio);
 717                         PIPE_LOCK(rpipe);
 718                         if (error)
 719                                 break;
 720
 721                         rpipe->pipe_buffer.out += size;
 722                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 723                                 rpipe->pipe_buffer.out = 0;
 724
 725                         rpipe->pipe_buffer.cnt -= size;
 726
 727                         /*
 728                          * If there is no more to read in the pipe, reset
 729                          * its pointers to the beginning.  This improves
 730                          * cache hit stats.
 731                          */
 732                         if (rpipe->pipe_buffer.cnt == 0) {
 733                                 rpipe->pipe_buffer.in = 0;
 734                                 rpipe->pipe_buffer.out = 0;
 735                         }
 736                         nread += size;
 737 #ifndef PIPE_NODIRECT
 738                 /*
 739                  * Direct copy, bypassing a kernel buffer.
 740                  */
 741                 } else if ((size = rpipe->pipe_map.cnt) &&
 742                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 743                         if (size > uio->uio_resid)
 744                                 size = (u_int) uio->uio_resid;
 745
 746                         PIPE_UNLOCK(rpipe);
 747                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 748                             rpipe->pipe_map.pos, size, uio);
 749                         PIPE_LOCK(rpipe);
 750                         if (error)
 751                                 break;
 752                         nread += size;
 753                         rpipe->pipe_map.pos += size;
 754                         rpipe->pipe_map.cnt -= size;
 755                         if (rpipe->pipe_map.cnt == 0) {
 756                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 757                                 wakeup(rpipe);
 758                         }
 759 #endif
 760                 } else {
 761                         /*
 762                          * detect EOF condition
 763                          * read returns 0 on EOF, no need to set error
 764                          */
 765                         if (rpipe->pipe_state & PIPE_EOF)
 766                                 break;
 767
 768                         /*
 769                          * If the "write-side" has been blocked, wake it up now.
 770                          */
 771                         if (rpipe->pipe_state & PIPE_WANTW) {
 772                                 rpipe->pipe_state &= ~PIPE_WANTW;
 773                                 wakeup(rpipe);
 774                         }
 775
 776                         /*
 777                          * Break if some data was read.
 778                          */
 779                         if (nread > 0)
 780                                 break;
 781
 782                         /*
 783                          * Unlock the pipe buffer for our remaining processing.
 784                          * We will either break out with an error or we will
 785                          * sleep and relock to loop.
 786                          */
 787                         pipeunlock(rpipe);
 788
 789                         /*
 790                          * Handle non-blocking mode operation or
 791                          * wait for more data.
 792                          */
 793                         if (fp->f_flag & FNONBLOCK) {
 794                                 error = EAGAIN;
 795                         } else {
 796                                 rpipe->pipe_state |= PIPE_WANTR;
 797                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 798                                     PRIBIO | PCATCH,
 799                                     "piperd", 0)) == 0)
 800                                         error = pipelock(rpipe, 1);
 801                         }
 802                         if (error)
 803                                 goto unlocked_error;
 804                 }
 805         }
 806 #ifdef MAC
 807 locked_error:
 808 #endif
 809         pipeunlock(rpipe);
 810
 811         /* XXX: should probably do this before getting any locks. */
 812         if (error == 0)
 813                 vfs_timestamp(&rpipe->pipe_atime);
 814 unlocked_error:
 815         --rpipe->pipe_busy;
 816
 817         /*
 818          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 819          */
 820         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 821                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 822                 wakeup(rpipe);
 823         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 824                 /*
 825                  * Handle write blocking hysteresis.
 826                  */
 827                 if (rpipe->pipe_state & PIPE_WANTW) {
 828                         rpipe->pipe_state &= ~PIPE_WANTW;
 829                         wakeup(rpipe);
 830                 }
 831         }
 832
 833         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 834                 pipeselwakeup(rpipe);
 835
 836         PIPE_UNLOCK(rpipe);
 837         return (error);
 838 }
 839
 840 #ifndef PIPE_NODIRECT
 841 /*
 842  * Map the sending processes' buffer into kernel space and wire it.
 843  * This is similar to a physical write operation.
 844  */
 845 static int
 846 pipe_build_write_buffer(wpipe, uio)
 847         struct pipe *wpipe;
 848         struct uio *uio;
 849 {
 850         u_int size;
 851         int i;
 852
 853         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 854         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 855                 ("Clone attempt on non-direct write pipe!"));
 856
 857         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 858                 size = wpipe->pipe_buffer.size;
 859         else
 860                 size = uio->uio_iov->iov_len;
 861
 862         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 863             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 864             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 865                 return (EFAULT);
 866
 867 /*
 868  * set up the control block
 869  */
 870         wpipe->pipe_map.npages = i;
 871         wpipe->pipe_map.pos =
 872             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 873         wpipe->pipe_map.cnt = size;
 874
 875 /*
 876  * and update the uio data
 877  */
 878
 879         uio->uio_iov->iov_len -= size;
 880         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 881         if (uio->uio_iov->iov_len == 0)
 882                 uio->uio_iov++;
 883         uio->uio_resid -= size;
 884         uio->uio_offset += size;
 885         return (0);
 886 }
 887
 888 /*
 889  * unmap and unwire the process buffer
 890  */
 891 static void
 892 pipe_destroy_write_buffer(wpipe)
 893         struct pipe *wpipe;
 894 {
 895
 896         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 897         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 898         wpipe->pipe_map.npages = 0;
 899 }
 900
 901 /*
 902  * In the case of a signal, the writing process might go away.  This
 903  * code copies the data into the circular buffer so that the source
 904  * pages can be freed without loss of data.
 905  */
 906 static void
 907 pipe_clone_write_buffer(wpipe)
 908         struct pipe *wpipe;
 909 {
 910         struct uio uio;
 911         struct iovec iov;
 912         int size;
 913         int pos;
 914
 915         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 916         size = wpipe->pipe_map.cnt;
 917         pos = wpipe->pipe_map.pos;
 918
 919         wpipe->pipe_buffer.in = size;
 920         wpipe->pipe_buffer.out = 0;
 921         wpipe->pipe_buffer.cnt = size;
 922         wpipe->pipe_state &= ~PIPE_DIRECTW;
 923
 924         PIPE_UNLOCK(wpipe);
 925         iov.iov_base = wpipe->pipe_buffer.buffer;
 926         iov.iov_len = size;
 927         uio.uio_iov = &iov;
 928         uio.uio_iovcnt = 1;
 929         uio.uio_offset = 0;
 930         uio.uio_resid = size;
 931         uio.uio_segflg = UIO_SYSSPACE;
 932         uio.uio_rw = UIO_READ;
 933         uio.uio_td = curthread;
 934         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 935         PIPE_LOCK(wpipe);
 936         pipe_destroy_write_buffer(wpipe);
 937 }
 938
 939 /*
 940  * This implements the pipe buffer write mechanism.  Note that only
 941  * a direct write OR a normal pipe write can be pending at any given time.
 942  * If there are any characters in the pipe buffer, the direct write will
 943  * be deferred until the receiving process grabs all of the bytes from
 944  * the pipe buffer.  Then the direct mapping write is set-up.
 945  */
 946 static int
 947 pipe_direct_write(wpipe, uio)
 948         struct pipe *wpipe;
 949         struct uio *uio;
 950 {
 951         int error;
 952
 953 retry:
 954         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 955         error = pipelock(wpipe, 1);
 956         if (wpipe->pipe_state & PIPE_EOF)
 957                 error = EPIPE;
 958         if (error) {
 959                 pipeunlock(wpipe);
 960                 goto error1;
 961         }
 962         while (wpipe->pipe_state & PIPE_DIRECTW) {
 963                 if (wpipe->pipe_state & PIPE_WANTR) {
 964                         wpipe->pipe_state &= ~PIPE_WANTR;
 965                         wakeup(wpipe);
 966                 }
 967                 pipeselwakeup(wpipe);
 968                 wpipe->pipe_state |= PIPE_WANTW;
 969                 pipeunlock(wpipe);
 970                 error = msleep(wpipe, PIPE_MTX(wpipe),
 971                     PRIBIO | PCATCH, "pipdww", 0);
 972                 if (error)
 973                         goto error1;
 974                 else
 975                         goto retry;
 976         }
 977         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 978         if (wpipe->pipe_buffer.cnt > 0) {
 979                 if (wpipe->pipe_state & PIPE_WANTR) {
 980                         wpipe->pipe_state &= ~PIPE_WANTR;
 981                         wakeup(wpipe);
 982                 }
 983                 pipeselwakeup(wpipe);
 984                 wpipe->pipe_state |= PIPE_WANTW;
 985                 pipeunlock(wpipe);
 986                 error = msleep(wpipe, PIPE_MTX(wpipe),
 987                     PRIBIO | PCATCH, "pipdwc", 0);
 988                 if (error)
 989                         goto error1;
 990                 else
 991                         goto retry;
 992         }
 993
 994         wpipe->pipe_state |= PIPE_DIRECTW;
 995
 996         PIPE_UNLOCK(wpipe);
 997         error = pipe_build_write_buffer(wpipe, uio);
 998         PIPE_LOCK(wpipe);
 999         if (error) {
1000                 wpipe->pipe_state &= ~PIPE_DIRECTW;
1001                 pipeunlock(wpipe);
1002                 goto error1;
1003         }
1004
1005         error = 0;
1006         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1007                 if (wpipe->pipe_state & PIPE_EOF) {
1008                         pipe_destroy_write_buffer(wpipe);
1009                         pipeselwakeup(wpipe);
1010                         pipeunlock(wpipe);
1011                         error = EPIPE;
1012                         goto error1;
1013                 }
1014                 if (wpipe->pipe_state & PIPE_WANTR) {
1015                         wpipe->pipe_state &= ~PIPE_WANTR;
1016                         wakeup(wpipe);
1017                 }
1018                 pipeselwakeup(wpipe);
1019                 wpipe->pipe_state |= PIPE_WANTW;
1020                 pipeunlock(wpipe);
1021                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1022                     "pipdwt", 0);
1023                 pipelock(wpipe, 0);
1024         }
1025
1026         if (wpipe->pipe_state & PIPE_EOF)
1027                 error = EPIPE;
1028         if (wpipe->pipe_state & PIPE_DIRECTW) {
1029                 /*
1030                  * this bit of trickery substitutes a kernel buffer for
1031                  * the process that might be going away.
1032                  */
1033                 pipe_clone_write_buffer(wpipe);
1034         } else {
1035                 pipe_destroy_write_buffer(wpipe);
1036         }
1037         pipeunlock(wpipe);
1038         return (error);
1039
1040 error1:
1041         wakeup(wpipe);
1042         return (error);
1043 }
1044 #endif
1045
1046 static int
1047 pipe_write(fp, uio, active_cred, flags, td)
1048         struct file *fp;
1049         struct uio *uio;
1050         struct ucred *active_cred;
1051         struct thread *td;
1052         int flags;
1053 {
1054         int error = 0;
1055         int desiredsize;
1056         ssize_t orig_resid;
1057         struct pipe *wpipe, *rpipe;
1058
1059         rpipe = fp->f_data;
1060         wpipe = PIPE_PEER(rpipe);
1061         PIPE_LOCK(rpipe);
1062         error = pipelock(wpipe, 1);
1063         if (error) {
1064                 PIPE_UNLOCK(rpipe);
1065                 return (error);
1066         }
1067         /*
1068          * detect loss of pipe read side, issue SIGPIPE if lost.
1069          */
1070         if (wpipe->pipe_present != PIPE_ACTIVE ||
1071             (wpipe->pipe_state & PIPE_EOF)) {
1072                 pipeunlock(wpipe);
1073                 PIPE_UNLOCK(rpipe);
1074                 return (EPIPE);
1075         }
1076 #ifdef MAC
1077         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1078         if (error) {
1079                 pipeunlock(wpipe);
1080                 PIPE_UNLOCK(rpipe);
1081                 return (error);
1082         }
1083 #endif
1084         ++wpipe->pipe_busy;
1085
1086         /* Choose a larger size if it's advantageous */
1087         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1088         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1089                 if (piperesizeallowed != 1)
1090                         break;
1091                 if (amountpipekva > maxpipekva / 2)
1092                         break;
1093                 if (desiredsize == BIG_PIPE_SIZE)
1094                         break;
1095                 desiredsize = desiredsize * 2;
1096         }
1097
1098         /* Choose a smaller size if we're in a OOM situation */
1099         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1100                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1101                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1102                 (piperesizeallowed == 1))
1103                 desiredsize = SMALL_PIPE_SIZE;
1104
1105         /* Resize if the above determined that a new size was necessary */
1106         if ((desiredsize != wpipe->pipe_buffer.size) &&
1107                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1108                 PIPE_UNLOCK(wpipe);
1109                 pipespace(wpipe, desiredsize);
1110                 PIPE_LOCK(wpipe);
1111         }
1112         if (wpipe->pipe_buffer.size == 0) {
1113                 /*
1114                  * This can only happen for reverse direction use of pipes
1115                  * in a complete OOM situation.
1116                  */
1117                 error = ENOMEM;
1118                 --wpipe->pipe_busy;
1119                 pipeunlock(wpipe);
1120                 PIPE_UNLOCK(wpipe);
1121                 return (error);
1122         }
1123
1124         pipeunlock(wpipe);
1125
1126         orig_resid = uio->uio_resid;
1127
1128         while (uio->uio_resid) {
1129                 int space;
1130
1131                 pipelock(wpipe, 0);
1132                 if (wpipe->pipe_state & PIPE_EOF) {
1133                         pipeunlock(wpipe);
1134                         error = EPIPE;
1135                         break;
1136                 }
1137 #ifndef PIPE_NODIRECT
1138                 /*
1139                  * If the transfer is large, we can gain performance if
1140                  * we do process-to-process copies directly.
1141                  * If the write is non-blocking, we don't use the
1142                  * direct write mechanism.
1143                  *
1144                  * The direct write mechanism will detect the reader going
1145                  * away on us.
1146                  */
1147                 if (uio->uio_segflg == UIO_USERSPACE &&
1148                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1149                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1150                     (fp->f_flag & FNONBLOCK) == 0) {
1151                         pipeunlock(wpipe);
1152                         error = pipe_direct_write(wpipe, uio);
1153                         if (error)
1154                                 break;
1155                         continue;
1156                 }
1157 #endif
1158
1159                 /*
1160                  * Pipe buffered writes cannot be coincidental with
1161                  * direct writes.  We wait until the currently executing
1162                  * direct write is completed before we start filling the
1163                  * pipe buffer.  We break out if a signal occurs or the
1164                  * reader goes away.
1165                  */
1166                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1167                         if (wpipe->pipe_state & PIPE_WANTR) {
1168                                 wpipe->pipe_state &= ~PIPE_WANTR;
1169                                 wakeup(wpipe);
1170                         }
1171                         pipeselwakeup(wpipe);
1172                         wpipe->pipe_state |= PIPE_WANTW;
1173                         pipeunlock(wpipe);
1174                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1175                             "pipbww", 0);
1176                         if (error)
1177                                 break;
1178                         else
1179                                 continue;
1180                 }
1181
1182                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1183
1184                 /* Writes of size <= PIPE_BUF must be atomic. */
1185                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1186                         space = 0;
1187
1188                 if (space > 0) {
1189                         int size;       /* Transfer size */
1190                         int segsize;    /* first segment to transfer */
1191
1192                         /*
1193                          * Transfer size is minimum of uio transfer
1194                          * and free space in pipe buffer.
1195                          */
1196                         if (space > uio->uio_resid)
1197                                 size = uio->uio_resid;
1198                         else
1199                                 size = space;
1200                         /*
1201                          * First segment to transfer is minimum of
1202                          * transfer size and contiguous space in
1203                          * pipe buffer.  If first segment to transfer
1204                          * is less than the transfer size, we've got
1205                          * a wraparound in the buffer.
1206                          */
1207                         segsize = wpipe->pipe_buffer.size -
1208                                 wpipe->pipe_buffer.in;
1209                         if (segsize > size)
1210                                 segsize = size;
1211
1212                         /* Transfer first segment */
1213
1214                         PIPE_UNLOCK(rpipe);
1215                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1216                                         segsize, uio);
1217                         PIPE_LOCK(rpipe);
1218
1219                         if (error == 0 && segsize < size) {
1220                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1221                                         wpipe->pipe_buffer.size,
1222                                         ("Pipe buffer wraparound disappeared"));
1223                                 /*
1224                                  * Transfer remaining part now, to
1225                                  * support atomic writes.  Wraparound
1226                                  * happened.
1227                                  */
1228
1229                                 PIPE_UNLOCK(rpipe);
1230                                 error = uiomove(
1231                                     &wpipe->pipe_buffer.buffer[0],
1232                                     size - segsize, uio);
1233                                 PIPE_LOCK(rpipe);
1234                         }
1235                         if (error == 0) {
1236                                 wpipe->pipe_buffer.in += size;
1237                                 if (wpipe->pipe_buffer.in >=
1238                                     wpipe->pipe_buffer.size) {
1239                                         KASSERT(wpipe->pipe_buffer.in ==
1240                                                 size - segsize +
1241                                                 wpipe->pipe_buffer.size,
1242                                                 ("Expected wraparound bad"));
1243                                         wpipe->pipe_buffer.in = size - segsize;
1244                                 }
1245
1246                                 wpipe->pipe_buffer.cnt += size;
1247                                 KASSERT(wpipe->pipe_buffer.cnt <=
1248                                         wpipe->pipe_buffer.size,
1249                                         ("Pipe buffer overflow"));
1250                         }
1251                         pipeunlock(wpipe);
1252                         if (error != 0)
1253                                 break;
1254                 } else {
1255                         /*
1256                          * If the "read-side" has been blocked, wake it up now.
1257                          */
1258                         if (wpipe->pipe_state & PIPE_WANTR) {
1259                                 wpipe->pipe_state &= ~PIPE_WANTR;
1260                                 wakeup(wpipe);
1261                         }
1262
1263                         /*
1264                          * don't block on non-blocking I/O
1265                          */
1266                         if (fp->f_flag & FNONBLOCK) {
1267                                 error = EAGAIN;
1268                                 pipeunlock(wpipe);
1269                                 break;
1270                         }
1271
1272                         /*
1273                          * We have no more space and have something to offer,
1274                          * wake up select/poll.
1275                          */
1276                         pipeselwakeup(wpipe);
1277
1278                         wpipe->pipe_state |= PIPE_WANTW;
1279                         pipeunlock(wpipe);
1280                         error = msleep(wpipe, PIPE_MTX(rpipe),
1281                             PRIBIO | PCATCH, "pipewr", 0);
1282                         if (error != 0)
1283                                 break;
1284                 }
1285         }
1286
1287         pipelock(wpipe, 0);
1288         --wpipe->pipe_busy;
1289
1290         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1291                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1292                 wakeup(wpipe);
1293         } else if (wpipe->pipe_buffer.cnt > 0) {
1294                 /*
1295                  * If we have put any characters in the buffer, we wake up
1296                  * the reader.
1297                  */
1298                 if (wpipe->pipe_state & PIPE_WANTR) {
1299                         wpipe->pipe_state &= ~PIPE_WANTR;
1300                         wakeup(wpipe);
1301                 }
1302         }
1303
1304         /*
1305          * Don't return EPIPE if I/O was successful
1306          */
1307         if ((wpipe->pipe_buffer.cnt == 0) &&
1308             (uio->uio_resid == 0) &&
1309             (error == EPIPE)) {
1310                 error = 0;
1311         }
1312
1313         if (error == 0)
1314                 vfs_timestamp(&wpipe->pipe_mtime);
1315
1316         /*
1317          * We have something to offer,
1318          * wake up select/poll.
1319          */
1320         if (wpipe->pipe_buffer.cnt)
1321                 pipeselwakeup(wpipe);
1322
1323         pipeunlock(wpipe);
1324         PIPE_UNLOCK(rpipe);
1325         return (error);
1326 }
1327
1328 /* ARGSUSED */
1329 static int
1330 pipe_truncate(fp, length, active_cred, td)
1331         struct file *fp;
1332         off_t length;
1333         struct ucred *active_cred;
1334         struct thread *td;
1335 {
1336
1337         /* For named pipes call the vnode operation. */
1338         if (fp->f_vnode != NULL)
1339                 return (vnops.fo_truncate(fp, length, active_cred, td));
1340         return (EINVAL);
1341 }
1342
1343 /*
1344  * we implement a very minimal set of ioctls for compatibility with sockets.
1345  */
1346 static int
1347 pipe_ioctl(fp, cmd, data, active_cred, td)
1348         struct file *fp;
1349         u_long cmd;
1350         void *data;
1351         struct ucred *active_cred;
1352         struct thread *td;
1353 {
1354         struct pipe *mpipe = fp->f_data;
1355         int error;
1356
1357         PIPE_LOCK(mpipe);
1358
1359 #ifdef MAC
1360         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1361         if (error) {
1362                 PIPE_UNLOCK(mpipe);
1363                 return (error);
1364         }
1365 #endif
1366
1367         error = 0;
1368         switch (cmd) {
1369
1370         case FIONBIO:
1371                 break;
1372
1373         case FIOASYNC:
1374                 if (*(int *)data) {
1375                         mpipe->pipe_state |= PIPE_ASYNC;
1376                 } else {
1377                         mpipe->pipe_state &= ~PIPE_ASYNC;
1378                 }
1379                 break;
1380
1381         case FIONREAD:
1382                 if (!(fp->f_flag & FREAD)) {
1383                         *(int *)data = 0;
1384                         PIPE_UNLOCK(mpipe);
1385                         return (0);
1386                 }
1387                 if (mpipe->pipe_state & PIPE_DIRECTW)
1388                         *(int *)data = mpipe->pipe_map.cnt;
1389                 else
1390                         *(int *)data = mpipe->pipe_buffer.cnt;
1391                 break;
1392
1393         case FIOSETOWN:
1394                 PIPE_UNLOCK(mpipe);
1395                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1396                 goto out_unlocked;
1397
1398         case FIOGETOWN:
1399                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1400                 break;
1401
1402         /* This is deprecated, FIOSETOWN should be used instead. */
1403         case TIOCSPGRP:
1404                 PIPE_UNLOCK(mpipe);
1405                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1406                 goto out_unlocked;
1407
1408         /* This is deprecated, FIOGETOWN should be used instead. */
1409         case TIOCGPGRP:
1410                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1411                 break;
1412
1413         default:
1414                 error = ENOTTY;
1415                 break;
1416         }
1417         PIPE_UNLOCK(mpipe);
1418 out_unlocked:
1419         return (error);
1420 }
1421
1422 static int
1423 pipe_poll(fp, events, active_cred, td)
1424         struct file *fp;
1425         int events;
1426         struct ucred *active_cred;
1427         struct thread *td;
1428 {
1429         struct pipe *rpipe;
1430         struct pipe *wpipe;
1431         int levents, revents;
1432 #ifdef MAC
1433         int error;
1434 #endif
1435
1436         revents = 0;
1437         rpipe = fp->f_data;
1438         wpipe = PIPE_PEER(rpipe);
1439         PIPE_LOCK(rpipe);
1440 #ifdef MAC
1441         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1442         if (error)
1443                 goto locked_error;
1444 #endif
1445         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1446                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1447                     (rpipe->pipe_buffer.cnt > 0))
1448                         revents |= events & (POLLIN | POLLRDNORM);
1449
1450         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1451                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1452                     (wpipe->pipe_state & PIPE_EOF) ||
1453                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1454                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1455                          wpipe->pipe_buffer.size == 0)))
1456                         revents |= events & (POLLOUT | POLLWRNORM);
1457
1458         levents = events &
1459             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1460         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1461             fp->f_seqcount == rpipe->pipe_wgen)
1462                 events |= POLLINIGNEOF;
1463
1464         if ((events & POLLINIGNEOF) == 0) {
1465                 if (rpipe->pipe_state & PIPE_EOF) {
1466                         revents |= (events & (POLLIN | POLLRDNORM));
1467                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1468                             (wpipe->pipe_state & PIPE_EOF))
1469                                 revents |= POLLHUP;
1470                 }
1471         }
1472
1473         if (revents == 0) {
1474                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1475                         selrecord(td, &rpipe->pipe_sel);
1476                         if (SEL_WAITING(&rpipe->pipe_sel))
1477                                 rpipe->pipe_state |= PIPE_SEL;
1478                 }
1479
1480                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1481                         selrecord(td, &wpipe->pipe_sel);
1482                         if (SEL_WAITING(&wpipe->pipe_sel))
1483                                 wpipe->pipe_state |= PIPE_SEL;
1484                 }
1485         }
1486 #ifdef MAC
1487 locked_error:
1488 #endif
1489         PIPE_UNLOCK(rpipe);
1490
1491         return (revents);
1492 }
1493
1494 /*
1495  * We shouldn't need locks here as we're doing a read and this should
1496  * be a natural race.
1497  */
1498 static int
1499 pipe_stat(fp, ub, active_cred, td)
1500         struct file *fp;
1501         struct stat *ub;
1502         struct ucred *active_cred;
1503         struct thread *td;
1504 {
1505         struct pipe *pipe;
1506         int new_unr;
1507 #ifdef MAC
1508         int error;
1509 #endif
1510
1511         pipe = fp->f_data;
1512         PIPE_LOCK(pipe);
1513 #ifdef MAC
1514         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1515         if (error) {
1516                 PIPE_UNLOCK(pipe);
1517                 return (error);
1518         }
1519 #endif
1520
1521         /* For named pipes ask the underlying filesystem. */
1522         if (pipe->pipe_state & PIPE_NAMED) {
1523                 PIPE_UNLOCK(pipe);
1524                 return (vnops.fo_stat(fp, ub, active_cred, td));
1525         }
1526
1527         /*
1528          * Lazily allocate an inode number for the pipe.  Most pipe
1529          * users do not call fstat(2) on the pipe, which means that
1530          * postponing the inode allocation until it is must be
1531          * returned to userland is useful.  If alloc_unr failed,
1532          * assign st_ino zero instead of returning an error.
1533          * Special pipe_ino values:
1534          *  -1 - not yet initialized;
1535          *  0  - alloc_unr failed, return 0 as st_ino forever.
1536          */
1537         if (pipe->pipe_ino == (ino_t)-1) {
1538                 new_unr = alloc_unr(pipeino_unr);
1539                 if (new_unr != -1)
1540                         pipe->pipe_ino = new_unr;
1541                 else
1542                         pipe->pipe_ino = 0;
1543         }
1544         PIPE_UNLOCK(pipe);
1545
1546         bzero(ub, sizeof(*ub));
1547         ub->st_mode = S_IFIFO;
1548         ub->st_blksize = PAGE_SIZE;
1549         if (pipe->pipe_state & PIPE_DIRECTW)
1550                 ub->st_size = pipe->pipe_map.cnt;
1551         else
1552                 ub->st_size = pipe->pipe_buffer.cnt;
1553         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1554         ub->st_atim = pipe->pipe_atime;
1555         ub->st_mtim = pipe->pipe_mtime;
1556         ub->st_ctim = pipe->pipe_ctime;
1557         ub->st_uid = fp->f_cred->cr_uid;
1558         ub->st_gid = fp->f_cred->cr_gid;
1559         ub->st_dev = pipedev_ino;
1560         ub->st_ino = pipe->pipe_ino;
1561         /*
1562          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1563          */
1564         return (0);
1565 }
1566
1567 /* ARGSUSED */
1568 static int
1569 pipe_close(fp, td)
1570         struct file *fp;
1571         struct thread *td;
1572 {
1573
1574         if (fp->f_vnode != NULL)
1575                 return vnops.fo_close(fp, td);
1576         fp->f_ops = &badfileops;
1577         pipe_dtor(fp->f_data);
1578         fp->f_data = NULL;
1579         return (0);
1580 }
1581
1582 static int
1583 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1584 {
1585         struct pipe *cpipe;
1586         int error;
1587
1588         cpipe = fp->f_data;
1589         if (cpipe->pipe_state & PIPE_NAMED)
1590                 error = vn_chmod(fp, mode, active_cred, td);
1591         else
1592                 error = invfo_chmod(fp, mode, active_cred, td);
1593         return (error);
1594 }
1595
1596 static int
1597 pipe_chown(fp, uid, gid, active_cred, td)
1598         struct file *fp;
1599         uid_t uid;
1600         gid_t gid;
1601         struct ucred *active_cred;
1602         struct thread *td;
1603 {
1604         struct pipe *cpipe;
1605         int error;
1606
1607         cpipe = fp->f_data;
1608         if (cpipe->pipe_state & PIPE_NAMED)
1609                 error = vn_chown(fp, uid, gid, active_cred, td);
1610         else
1611                 error = invfo_chown(fp, uid, gid, active_cred, td);
1612         return (error);
1613 }
1614
1615 static void
1616 pipe_free_kmem(cpipe)
1617         struct pipe *cpipe;
1618 {
1619
1620         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1621             ("pipe_free_kmem: pipe mutex locked"));
1622
1623         if (cpipe->pipe_buffer.buffer != NULL) {
1624                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1625                 vm_map_remove(pipe_map,
1626                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1627                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1628                 cpipe->pipe_buffer.buffer = NULL;
1629         }
1630 #ifndef PIPE_NODIRECT
1631         {
1632                 cpipe->pipe_map.cnt = 0;
1633                 cpipe->pipe_map.pos = 0;
1634                 cpipe->pipe_map.npages = 0;
1635         }
1636 #endif
1637 }
1638
1639 /*
1640  * shutdown the pipe
1641  */
1642 static void
1643 pipeclose(cpipe)
1644         struct pipe *cpipe;
1645 {
1646         struct pipepair *pp;
1647         struct pipe *ppipe;
1648
1649         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1650
1651         PIPE_LOCK(cpipe);
1652         pipelock(cpipe, 0);
1653         pp = cpipe->pipe_pair;
1654
1655         pipeselwakeup(cpipe);
1656
1657         /*
1658          * If the other side is blocked, wake it up saying that
1659          * we want to close it down.
1660          */
1661         cpipe->pipe_state |= PIPE_EOF;
1662         while (cpipe->pipe_busy) {
1663                 wakeup(cpipe);
1664                 cpipe->pipe_state |= PIPE_WANT;
1665                 pipeunlock(cpipe);
1666                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1667                 pipelock(cpipe, 0);
1668         }
1669
1670
1671         /*
1672          * Disconnect from peer, if any.
1673          */
1674         ppipe = cpipe->pipe_peer;
1675         if (ppipe->pipe_present == PIPE_ACTIVE) {
1676                 pipeselwakeup(ppipe);
1677
1678                 ppipe->pipe_state |= PIPE_EOF;
1679                 wakeup(ppipe);
1680                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1681         }
1682
1683         /*
1684          * Mark this endpoint as free.  Release kmem resources.  We
1685          * don't mark this endpoint as unused until we've finished
1686          * doing that, or the pipe might disappear out from under
1687          * us.
1688          */
1689         PIPE_UNLOCK(cpipe);
1690         pipe_free_kmem(cpipe);
1691         PIPE_LOCK(cpipe);
1692         cpipe->pipe_present = PIPE_CLOSING;
1693         pipeunlock(cpipe);
1694
1695         /*
1696          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1697          * PIPE_FINALIZED, that allows other end to free the
1698          * pipe_pair, only after the knotes are completely dismantled.
1699          */
1700         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1701         cpipe->pipe_present = PIPE_FINALIZED;
1702         seldrain(&cpipe->pipe_sel);
1703         knlist_destroy(&cpipe->pipe_sel.si_note);
1704
1705         /*
1706          * If both endpoints are now closed, release the memory for the
1707          * pipe pair.  If not, unlock.
1708          */
1709         if (ppipe->pipe_present == PIPE_FINALIZED) {
1710                 PIPE_UNLOCK(cpipe);
1711 #ifdef MAC
1712                 mac_pipe_destroy(pp);
1713 #endif
1714                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1715         } else
1716                 PIPE_UNLOCK(cpipe);
1717 }
1718
1719 /*ARGSUSED*/
1720 static int
1721 pipe_kqfilter(struct file *fp, struct knote *kn)
1722 {
1723         struct pipe *cpipe;
1724
1725         /*
1726          * If a filter is requested that is not supported by this file
1727          * descriptor, don't return an error, but also don't ever generate an
1728          * event.
1729          */
1730         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1731                 kn->kn_fop = &pipe_nfiltops;
1732                 return (0);
1733         }
1734         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1735                 kn->kn_fop = &pipe_nfiltops;
1736                 return (0);
1737         }
1738         cpipe = fp->f_data;
1739         PIPE_LOCK(cpipe);
1740         switch (kn->kn_filter) {
1741         case EVFILT_READ:
1742                 kn->kn_fop = &pipe_rfiltops;
1743                 break;
1744         case EVFILT_WRITE:
1745                 kn->kn_fop = &pipe_wfiltops;
1746                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1747                         /* other end of pipe has been closed */
1748                         PIPE_UNLOCK(cpipe);
1749                         return (EPIPE);
1750                 }
1751                 cpipe = PIPE_PEER(cpipe);
1752                 break;
1753         default:
1754                 PIPE_UNLOCK(cpipe);
1755                 return (EINVAL);
1756         }
1757
1758         kn->kn_hook = cpipe;
1759         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1760         PIPE_UNLOCK(cpipe);
1761         return (0);
1762 }
1763
1764 static void
1765 filt_pipedetach(struct knote *kn)
1766 {
1767         struct pipe *cpipe = kn->kn_hook;
1768
1769         PIPE_LOCK(cpipe);
1770         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1771         PIPE_UNLOCK(cpipe);
1772 }
1773
1774 /*ARGSUSED*/
1775 static int
1776 filt_piperead(struct knote *kn, long hint)
1777 {
1778         struct pipe *rpipe = kn->kn_hook;
1779         struct pipe *wpipe = rpipe->pipe_peer;
1780         int ret;
1781
1782         PIPE_LOCK(rpipe);
1783         kn->kn_data = rpipe->pipe_buffer.cnt;
1784         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1785                 kn->kn_data = rpipe->pipe_map.cnt;
1786
1787         if ((rpipe->pipe_state & PIPE_EOF) ||
1788             wpipe->pipe_present != PIPE_ACTIVE ||
1789             (wpipe->pipe_state & PIPE_EOF)) {
1790                 kn->kn_flags |= EV_EOF;
1791                 PIPE_UNLOCK(rpipe);
1792                 return (1);
1793         }
1794         ret = kn->kn_data > 0;
1795         PIPE_UNLOCK(rpipe);
1796         return ret;
1797 }
1798
1799 /*ARGSUSED*/
1800 static int
1801 filt_pipewrite(struct knote *kn, long hint)
1802 {
1803         struct pipe *wpipe;
1804
1805         wpipe = kn->kn_hook;
1806         PIPE_LOCK(wpipe);
1807         if (wpipe->pipe_present != PIPE_ACTIVE ||
1808             (wpipe->pipe_state & PIPE_EOF)) {
1809                 kn->kn_data = 0;
1810                 kn->kn_flags |= EV_EOF;
1811                 PIPE_UNLOCK(wpipe);
1812                 return (1);
1813         }
1814         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1815             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1816         if (wpipe->pipe_state & PIPE_DIRECTW)
1817                 kn->kn_data = 0;
1818
1819         PIPE_UNLOCK(wpipe);
1820         return (kn->kn_data >= PIPE_BUF);
1821 }
1822
1823 static void
1824 filt_pipedetach_notsup(struct knote *kn)
1825 {
1826
1827 }
1828
1829 static int
1830 filt_pipenotsup(struct knote *kn, long hint)
1831 {
1832
1833         return (0);
1834 }